SphereSfM Dataset (Dual-Res: low-res SfM + 8K faces)

If you've built a splat dataset from a panorama, you know the tension: structure-from-motion is expensive, but the training images are only as sharp as the frames you fed SfM. The dual-res variant of SphereSfM Dataset is the author's answer to "why not do both". It runs SfM on low-res equirect frames and reprojects the trainable cube faces from 8K panoramas on disk. You get cheap exhaustive matching and full-resolution training images, which is exactly the pipeline the main 1_generate-dataset-hires.json workflow uses.

The reasoning is sound and worth internalizing: SPHERE camera poses are angular, so they're effectively resolution-independent. Feature extraction, matching, and bundle adjustment don't need 8K - they need enough pixels to find corners. So this node spends the cheap frames on posing the scene and the expensive ones only where they actually matter: the pinhole faces your trainer trains on.

How it works

pano_frames_1..4 are the low-res equirect trajectories (say the raw 1440×720 WAN panoramas - a few GB for a couple hundred frames, so they fit fine in a ComfyUI tensor). The hi-res set is where it gets clever: hires_dir points at a folder of matching 8K equirects (e.g. <dataset>/panoramas_upscaled), read frame-by-frame from disk, never tensored - which is what avoids the OOM you'd hit trying to cram 122 GB of 8K into VRAM. After the low-res solve, the SPHERE camera is rescaled to the 8K grid and sphere_cubic_reprojecer samples the sharp source.

Two ways to feed the hi-res side:

• hires_dir + hires_glob - a folder of matching hi-res files. Sorted order and count must match the wired frames 1:1. Thin with this node's frame_stride, not the loader's, so both sides stay aligned.
• hires_1..hires_4 - wire a HiRes Composite's hires_manifest per trajectory. That carries the exact 8K file list and overrides hires_dir/hires_glob entirely.

Leave hires_dir empty and it degrades to a plain single-res SphereSfM run - same node, no special setup.

Inputs that matter

• pano_frames_1 (required) - the equirect trajectory SfM poses.
• output_name - dataset folder under ComfyUI/output or an absolute path.
• matcher_type - exhaustive is the default here (unlike the base node) because linking non-adjacent trajectories into one model is the whole point; sequential only matches temporally adjacent frames.
• on_split - stop (default) raises with a per-model breakdown if SfM forms more than one disconnected model, so you see that your trajectories didn't fuse instead of silently training on the biggest one. largest reprojects the biggest model anyway (legacy behavior).
• face_size - up to 8192 px; 0 = COLMAP default scaled from the rescaled 8K SPHERE camera (full detail).
• frame_stride / max_frames - thin the low-res frames and the matching hi-res files together, keeping them in lockstep.
• cleanup_hires_source - leave off unless you're done iterating. On, it deletes each consumed 8K file from its original folder only after a same-size copy is confirmed staged in the dataset's _spheresfm_work/equirect_hires. Worth knowing: if both are on the same volume they're hardlinks, so it frees a directory listing but ~no disk.

Outputs are the same trio as the base node: model_dir, num_images, num_points.

Install & gotchas

Installation is the pack-wide one - clone, pip install requirements.txt, restart, let colmap_sphere.exe auto-download on first run. You supply the WAN 2.1 i2v checkpoint and the converted Matrix-3D pano LoRA.

The two failure modes that bite people:

• Count mismatch. If the hi-res folder doesn't line up 1:1 with the low-res frames, the node errors. That's why striding happens here, not in an image loader - the loader can't thin both sides together.
• on_split fires. It means your trajectories didn't share enough view to fuse into one model. The fix is more overlap / more shared content between trajectories, or a lower init_min_tri_angle - not largest, which silently throws away everything outside the biggest model.

The underlying SfM rules still apply: real parallax, textured scene, and "No good initial image pair found" means lower init_min_tri_angle (the default here is already 4°, well below COLMAP's 16°). If your splats look soft, check the trainer's point cap - dual-res easily produces a few million points, and a low default cap (1M) will bottleneck the reconstruction and hide the resolution gain.